Self-adaption Wi-Fi tracking method based on environmental obstacle constraint

Abstract

The invention discloses a self-adaptive Wi-Fi tracking method based on environmental obstacle constraints, and belongs to the technical field of Wi-Fi tracking methods. According to the method, environmental obstacle constraints are adaptively considered in the tracking process, so that the generated trajectory is more in line with the actual situation, and the change of the environmental obstacle layout can be automatically coped with. According to the invention, the tracking precision in a complex and changeable scene is improved on the basis that the manpower cost does not need to be increased, so that the feasibility of deploying Wi-Fi equipment-free tracking in a real indoor scene is enhanced; the challenges that a traditional equipment-free tracking method neglects the limitation of an environment object on human body movement, and wide retraining or map reconstruction needs to be carried out when the environment layout changes are successfully solved.

Description

Technical Field

[0001] The present invention relates to the technical field of Wi-Fi tracking methods, and in particular to a Wi-Fi tracking method based on environmental obstacle constraint adaptation. Background Art

[0002] Device-free Wi-Fi tracking has become a hot topic in the field of indoor tracking. This innovative method does not require users to carry specific devices or sensors, which makes it show great potential in many emerging application scenarios, especially in the fields of elderly care and smart home. In addition, the seamless integration and privacy protection characteristics of this technology make it easier for users to be widely accepted and applied without worrying about the leakage of personal information.

[0003] Currently, most device-free Wi-Fi tracking systems work on the principle that the user's movement changes the propagation path of the signal, allowing location features to be extracted from the collected signal readings. Generally speaking, less signal noise means higher positioning accuracy.

[0004] Device-free Wi-Fi tracking implementations are generally divided into two categories: model-based methods and data-driven methods. Model-based methods extract signal features from raw Wi-Fi readings and then build models to infer the user's trajectory. However, it ignores the limitations of physical space, such as the impossibility of human movement passing through objects in the environment, which causes the prediction results to deviate from the actual situation, resulting in higher errors. Data-driven methods use deep learning techniques to integrate more information and improve positioning accuracy. However, when the environment layout changes, such as rearranging furniture, the universality of the neural network will be affected. Adapting to these changes usually requires re-collecting data or retraining the model, which reduces its practicality and scalability. In order to solve the above problems, the present invention proposes a Wi-Fi tracking method based on environmental obstacle constraint adaptation. Summary of the invention

[0005] 1. Technical Problems to be Solved by the Present Invention

[0006] The purpose of the present invention is to propose a Wi-Fi tracking method based on environmental obstacle constraint adaptation to effectively reduce the impact of noise caused by obstacles in complex indoor scenes on tracking. The present invention adaptively considers environmental obstacle constraints during the tracking process, so that the generated trajectory is more in line with the actual situation and can automatically respond to changes in the layout of environmental obstacles. The present invention improves the tracking accuracy in complex and changeable scenes without increasing labor costs, thereby enhancing the feasibility of deploying Wi-Fi device-free tracking in real indoor scenes.

[0007] 2. Technical solution

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A Wi-Fi tracking method based on environmental obstacle constraint adaptation is proposed. The method adaptively considers environmental obstacle constraints during the tracking process, so that the generated trajectory conforms to the actual situation and automatically responds to changes in the layout of environmental obstacles. The specific implementation method is as follows:

[0010] Build a constrained adaptive model: Design a device-free Wi-Fi tracking model that incorporates environmental information, takes environmental obstacle features as input, and generates fused features for different environmental characteristics while being tightly integrated with the tracking process;

[0011] Build an automatic map update mechanism: obtain real-time and accurate environmental information to improve tracking accuracy, maintain and update the input map using grayscale images generated from historical trajectory data, and automatically adapt to the changing environment layout;

[0012] Perform constraint-enhanced training: select trajectories near obstacles by filtering simulation data to improve the model's ability to learn constraint information; during the training phase, enhance the ability to learn constraints through data filtering and sequential freezing modules.

[0013] Preferably, the constraint adaptive model includes a PLCR extractor, a map encoder, a fusion module and a decision module;

[0014] The PLCR extractor is used to quickly retrieve preliminary trajectory features from the PLCR map. The method of obtaining PLCR using the PLCR extractor includes the following steps: first, cleaning the CSI data based on the CSI ratio and various filtering techniques; then, applying the Short-Time Fourier Transform (STFT) to determine the time distribution of the DFS and calculate the related PLCR; finally, using the Long Short-Term Memory (LSTM) to extract the PLCR feature sequence and generate the initial trajectory feature vector;

[0015] The map encoder is used to identify the distribution and position of obstacles in the environment. First, an obstacle map is constructed as an input map, and then an encoder is designed to divide the constructed obstacle map into multiple sub-blocks, and the distribution of 0 pixel values ​​in the image of each sub-block area is extracted respectively to obtain obstacle features;

[0016] The fusion module is used to integrate the map features into the preliminary trajectory features; the fusion module expands the map features by adding an extra dimension that matches the sequence length in the initial trajectory features, and connects the two along the extra dimension to obtain a combined feature; and further uses a gated recurrent unit (GRU) network to process the obtained combined feature, and dynamically refines the trajectory features using obstacle information to obtain rich fused features;

[0017] The decision module is used to map the integrated features to the output trajectory, use a fully connected layer to map the fused features, and generate the final two-dimensional predicted trajectory coordinates; the decision module captures the influence of the fused features on the predicted trajectory coordinates, and helps the constraint adaptive model learn the inaccessible areas caused by obstacles in the output space.

[0018] Preferably, the automatic map update mechanism comprises the following contents:

[0019] Construct a probability grayscale map: introduce a probability matrix and a grayscale map, and represent the existence of obstacles in the environment area as a probability value. The size of the probability value is proportional to the possibility of the obstacle appearing at that location. Each value in the probability matrix corresponds to the probability of the existence of an obstacle in the real environment.

[0020] Update the probability matrix: Use the user behavior history to estimate the probability of an obstacle at each point in the area and complete the update of the probability matrix;

[0021] Update dynamic grayscale map: Use the locally stored probability matrix to dynamically update the locally stored input map, including:

[0022] 1) Initialize the input map according to environmental conditions;

[0023] 2) Generate predicted trajectories from the collected user data;

[0024] 3) Update the probability matrix based on the predicted trajectory to determine whether the input map needs to be updated.

[0025] Preferably, the objectives of the restriction enhancement training are:

[0026] 1) Minimize the distance between the predicted trajectory and the target trajectory;

[0027] 2) Minimize the intersection between the predicted trajectory and the obstacle;

[0028] The training goal realization process includes the following:

[0029] The mean square error is used to measure the distance between the predicted trajectory and the target trajectory:

[0030]

[0031] Where n represents the number of points on the target trajectory, (x i ,y i ) represents a point on the target trajectory, represents a point on the predicted trajectory;

[0032] Define an additional loss function to quantitatively evaluate the intersection between the predicted trajectory and obstacles to ensure that the loss function also supports gradient propagation; use the F.grid_sample function in the PyTorch framework to query and extract pixel values ​​from the grid specified position; extract pixel values ​​from the input obstacle map corresponding to each predicted trajectory point, and quantify the degree of intersection between the predicted trajectory and the obstacle by calculating the difference between the pixel value and the 1-pixel map, and then obtain the obstacle loss. The specific function is expressed as:

[0033]

[0034] Among them, L OC Indicates the obstacle loss, Represents the pixel value in the input obstacle map corresponding to the predicted trajectory point.

[0035] Preferably, the strategy of limiting enhanced training is:

[0036] The PLCR extractor, map encoder, fusion module and decision module are divided into two parts for modular training. The specific contents are as follows:

[0037] The first part is the tracking network, which includes the PLCR extractor and decision model, and is used to implement basic tracking functions;

[0038] The second part is the map network, which consists of a map extractor and a fusion module, used to extract environmental information as auxiliary features;

[0039] The specific training process is as follows:

[0040] 1) Establish the basic tracking effect of the tracking network;

[0041] 2) Initialize the map network to ensure its compatibility with the tracking network;

[0042] 3) screening the training data to select cases where the predicted trajectory intersects with obstacles;

[0043] 4) Using the screened data, combined with the mean square error and obstacle error, a fusion loss is designed to perform adversarial training on the map network.

[0044] 3. Beneficial effects

[0045] (1) The present invention uses a neural network to learn the constraints imposed by obstacles on the motion trajectory, integrates environmental information and adaptively uses environmental information for tracking, and uses the distribution of objects in the environment to correct trajectories that are inconsistent with physical logic, thereby significantly reducing tracking errors.

[0046] (2) The present invention introduces a grayscale map to represent the probability of obstacle presence, which is updated with historical trajectory data. The method can run continuously in a changing environment, automatically detect changes in obstacle layout, and dynamically update and maintain the obstacle map to adapt to the new configuration without retraining or relabeling.

[0047] (3) Through a large number of experiments, the present invention verifies that the proposed Wi-Fi tracking technology based on environmental obstacle constraint adaptation can reduce the tracking error by up to 49.48% in various complex environments, and successfully deploys the model on embedded devices with low computing power and low memory, meeting the requirements of low latency and low energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A method flow chart of a Wi-Fi tracking method based on environmental obstacle constraint adaptation proposed by the present invention;

[0049] Figure 2 This is a flowchart of the map construction method and obstacle feature extraction mentioned in Example 1 of the present invention;

[0050] Figure 3 This is a schematic diagram of the dynamic grayscale map self-update mechanism mentioned in Example 1 of the present invention;

[0051] Figure 4 A comparison chart of the prediction effect (a) of the benchmark model mentioned in Example 1 of the present invention and the prediction effect (b) of the present invention;

[0052] Figure 5 The overall error comparison statistical graph (a) of scenario 1 and the overall error comparison statistical graph (b) of scenario 2 mentioned in Example 1 of the present invention are shown. DETAILED DESCRIPTION

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0054] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0056] The present invention proposes a device-free Wi-Fi tracking method that dynamically integrates environmental information, represents objects that hinder the user's movement in an indoor environment (such as furniture and walls) as obstacles, records the layout of obstacles in the environment as environmental information, and obstacle constraints indicate that the user's movement trajectory cannot pass through these objects. The environmental obstacle constraint adaptive Wi-Fi tracking framework designed by the present invention designs a lightweight dual-branch model to effectively integrate environmental information with Wi-Fi signal characteristics. At the same time, the present invention proposes an automatic map update mechanism that enables the system to adapt to changes in environmental layout without human intervention. Finally, a data-driven phased training strategy is implemented to improve the model's ability to learn constraints. The present invention successfully solves the challenges of traditional device-free tracking methods that ignore the restrictions of environmental objects on human movement and require extensive retraining or map reconstruction when the environmental layout changes.

[0057] Terminology explanation:

[0058] Wi-Fi passive positioning is a technology that determines the user's location based on analyzing Wi-Fi signals reflected by the human body. Unlike active positioning, Wi-Fi passive positioning does not require users to carry additional equipment, such as sensors, but only achieves positioning by analyzing the physical characteristics of the user's reflected signals.

[0059] Channel State Information (CSI) is a type of information used to represent the channel status in wireless communications. CSI includes data such as the amplitude, frequency, and phase of the signal, reflecting the influence of factors such as power attenuation and scattering on the signal during propagation.

[0060] Recurrent Neural Network (RNN) is a neural network architecture specifically designed for processing sequence data. RNN establishes connections between nodes and introduces data correlation in the time dimension, so that the model has a certain memory capacity and the ability to process time series information.

[0061] The path length change rate (PLCR) refers to the rate at which the path length of a signal changes after it is reflected from the human body surface. It is one of the important factors causing Doppler frequency shift (DFS). According to the Fresnel zone theory, by analyzing the PLCR, the normal velocity on a certain signal link can be obtained. After the normal velocities of multiple links are synthesized, the actual speed and direction of the user can be obtained, thereby realizing the positioning function.

[0062] Feature fusion refers to the integration of information from different sources or different feature sets during data processing and analysis to improve the performance and accuracy of the model. Through feature fusion, the advantages of each feature can be combined and the information of multiple features can be integrated to obtain a more comprehensive and effective data representation.

[0063] Environmental Obstacles Constraint refers to objects in the environment, such as furniture and walls, that hinder the user's movement, limit the user's range of activities, and make certain areas inaccessible, forming physical space constraints imposed by obstacles. The existence of this constraint prevents human movement trajectories from passing through obstacles, and stipulates that the movement trajectory must meet physical logic.

[0064] Based on the above content, a Wi-Fi tracking method based on environmental obstacle constraint adaptation proposed by the present invention is described below in conjunction with a specific example, and the specific content is as follows.

[0065] Embodiment 1:

[0066] The present invention uses a lightweight model to fuse environmental information and Wi-Fi features, realizes the correction of predicted trajectories using environmental information, and proposes a probability grayscale map updated based on historical trajectory data, establishes an automatic update mechanism, allows the model to effectively adapt to new layouts, avoids manual map reconstruction, and adopts a data-driven strategy to accurately simulate environmental constraints by generating data reflecting real-world obstacles, introduces obstacle loss to minimize trajectory errors and ensure the minimum intersection with obstacles. The specific implementation method includes the following three parts:

[0067] Build a constraint adaptive model:

[0068] In order to utilize environmental information to improve tracking accuracy and enable the model to handle different obstacle distributions, the present invention innovatively uses environmental features as the input of the model, thereby constructing a device-free tracking model for Wi-Fi based on adaptive learning obstacle constraints. The input of the model includes two components: PLCR features from Wi-Fi data and image features based on the layout of environmental obstacles. The model uses the Wi-Fi modality as the main modality and incorporates environmental information while maintaining the original tracking performance. In addition, by utilizing images as an auxiliary source of tracking information, the restrictions on user movement caused by environmental obstacles are incorporated into the tracking model. This method integrates these two types of heterogeneous signals, ensuring that the model can effectively adapt to various obstacle layouts.

[0069] The model mainly consists of four modules: a PLCR extractor that quickly retrieves preliminary trajectory features from the PLCR map; a map encoder that identifies the distribution and location of obstacles in the environment; a fusion module that integrates map features into preliminary trajectory features; and a decision module that maps the integrated features to the output trajectory.

[0070] PLCR Extractor: User movement changes the reflection path length, causing the received signal frequency to change accordingly, which is called Doppler frequency shift (DFS). In order to achieve device-free tracking, it is necessary to analyze the DFS and the corresponding PLCR, and convert the PLCR into the actual user speed. The method for obtaining PLCR includes the following steps: first, the CSI data is cleaned based on the CSI ratio and various filtering techniques; then, the short-time Fourier transform (STFT) is applied to determine the time distribution of DFS and calculate the related PLCR. Finally, the long short-term memory (LSTM) is used to extract the PLCR feature sequence and generate the initial trajectory feature vector.

[0071] Map encoder: In order to extract the obstacle features required for learning constraints from the environment, we first need to build an input map and then design an encoder to extract obstacle features from the image. In practical applications, there are various effective methods to obtain a planar top view of the environment obstacle layout. Figure 2As shown, first, a proportional blank two-dimensional schematic diagram (called a blank map) is created according to the actual area of ​​the environment, and all pixel values ​​are set to 1. According to the view layout, the pixels corresponding to the area occupied by the obstacles are marked as 0 on the blank map. Finally, all pixel values ​​are merged into a black and white image, namely the obstacle map, as the input of the model. This labeling method can clearly represent the spatial distribution of obstacles in the environment and avoid cumbersome environmental space description. The present invention designs a map encoder, which divides the obstacle map into multiple sub-blocks and extracts the distribution of 0 pixel values ​​in each sub-area image to obtain obstacle features.

[0072] Fusion module: The initial trajectory feature is a three-dimensional vector of variable size (B, L, H), where B corresponds to the batch size, L represents the sequence length (variable length, depending on the duration of actual data collection), and H represents the number of features at each time step. The map feature is represented as a two-dimensional vector of fixed length, denoted as (B, R), where B represents the batch size and R represents the number of features. The present invention expands the map feature by adding an additional dimension that matches the size of L in the initial trajectory feature, and connects the two along this dimension to obtain a combined feature. In order to utilize environmental information to achieve coordinate-level correction, the module uses a gated recurrent unit (GRU) network to process the connected combined features, so that the trajectory features can be dynamically refined using obstacle information to obtain rich fused features.

[0073] Decision module: This module uses a fully connected layer to map the fused features and generate the final 2D predicted trajectory coordinates. The fully connected layer effectively captures the impact of the fused features on the predicted trajectory coordinates and helps the model learn the inaccessible areas caused by obstacles in the output space.

[0074] Build an automatic map update mechanism:

[0075] In practical applications, the initial distribution of obstacles may be difficult to determine, or may change dynamically over time, so map information needs to be updated regularly. Although the manual update process is not complicated, the frequent update requirements make it cumbersome and inefficient. Therefore, the present invention proposes a probabilistic grayscale map concept and an automatic map update mechanism to solve the situation where the initial environment layout is unknown or the layout of environmental objects changes frequently. This method eliminates the need for secondary annotation of the input map and additional training of the model, thereby improving the flexibility and efficiency of the system in practical applications.

[0076] Probabilistic grayscale map: In environments where the initial position of obstacles is unknown or changes frequently, re-marking map input may be delayed or inaccurate. When the static binary obstacle feature map does not reflect the current state, significant errors may occur, especially when the new trajectory intersects with the area previously occupied by the obstacle. To solve this problem, the present invention introduces a probability matrix and a grayscale map. Specifically, the presence of obstacles in the environmental area is represented as a probability value, where the higher the value, the greater the possibility of an obstacle appearing at that location. Each value in the probability matrix corresponds to the probability of an obstacle existing in the real environment.

[0077] User behavior update probability matrix: The present invention uses historical test trajectories to estimate the probability of obstacles at each point in the area. Since the probability of obstacles appearing in the area covered by the historical trajectories is low, the probability of the areas traversed by these user behaviors can be reduced during the update process.

[0078] Dynamic grayscale map update: This paper proposes a mechanism to update the input map based on historical test trajectories. This mechanism establishes an automatic update framework to dynamically update the locally stored input map using the locally stored probability matrix, such as Figure 3 It consists of three parts: 1) Initializing the input map according to environmental conditions; 2) Generating predicted trajectories from collected user data; 3) Updating the probability matrix based on the predicted trajectory to determine whether the input map needs to be updated. This framework ensures the timeliness and accuracy of the input map and provides a robust data foundation for trajectory correction using environmental information. Figure 1 The input map is then stored locally, thus avoiding the need for the user to input map information when predicting each trajectory. The received PLCR features are combined with the locally stored map to generate the current predicted user trajectory, thereby updating the probability matrix. When the map needs to be updated, a grayscale map is generated based on the updated probability matrix to reduce potential differences in trajectory predictions due to outdated map information. Finally, the grayscale map replaces the input map, realizing automatic updating of map information. This mechanism allows the input map to be automatically updated without manual reconstruction.

[0079] Limiting plyometric training:

[0080] After determining the model architecture and workflow, the model is trained to adapt it to the task. During the model training process, two main goals are expected to be achieved: 1) minimize the distance between the predicted trajectory and the target trajectory 2) minimize the intersection between the predicted trajectory and the obstacle.

[0081] Obstacle loss: Existing methods use mean squared error (MSE) to measure the distance between the predicted trajectory and the label:

[0082]

[0083] Where n represents the number of points on the trajectory, (x i ,y i ) represents a point on the marker trajectory, represents a point on the predicted trajectory. However, L MSE Without considering the inaccessible areas caused by obstacles, it cannot fully meet the training requirements. Therefore, an additional loss function must be defined to quantitatively evaluate the intersection between the predicted trajectory and obstacles to ensure that this loss function also supports gradient propagation. By utilizing the F.grid_sample function in the PyTorch framework (which accepts an input tensor and a grid tensor), it is possible to efficiently query and extract pixel values ​​from the locations specified by the grid. Using this function, extract pixel values ​​from the input obstacle map corresponding to each predicted trajectory point The difference between these pixel values ​​and the 1-pixel map is calculated to quantify the degree of intersection between the predicted trajectory and the obstacle, thereby obtaining the obstacle loss L OC :

[0084]

[0085] Model training strategy: The model designed by the present invention has a dual-branch structure with Wi-Fi features as the main form. In addition, since the model needs to learn how to use obstacle restrictions to correct trajectories and reduce errors while ensuring basic tracking accuracy, the present invention divides the four modules into two parts for modular training: the first is the tracking network, including the PLCR extractor and the decision model, to achieve basic tracking functions; the second is the map network, which consists of a map extractor and a fusion module to extract environmental information as an auxiliary feature. The training process is divided into four stages: in the first stage, the basic tracking effect of the tracking network is established; in the second stage, the map network is initialized to ensure its compatibility with the tracking network; in the third stage, the training data is screened to select situations where the predicted trajectory intersects with obstacles; in the fourth stage, adversarial training is performed on the map network using the screened data and the fusion loss combining the mean square error and the obstacle error. By gradually freezing the model training and data enhancement, the model's ability to use obstacle restrictions to correct trajectories and reduce errors is enhanced.

[0086] Comparative experimental verification:

[0087] In order to verify the trajectory correction effect and versatility of the present invention, two different environments (tracking ranges of 3m×6m and 6m×6m respectively) were selected for testing, and a network without obstacle restrictions was used as a comparison for the method proposed in the present invention, which was recorded as the benchmark model. Figure 4-5 As shown, Figure 4A comparison chart of the tracking effects of the same data under the baseline model (a) and the model of the present invention (b); Figure 5 The overall error comparison statistics of scene 1 (a) and scene 2 (b) are shown. Figure 4-5 It can be seen from the content that the trajectory predicted by the present invention can pass through fewer obstacles and achieve higher tracking accuracy.

[0088] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.

Claims

1. A Wi-Fi tracking method based on environmental obstacle constraint adaptation, characterized in that: The method adaptively considers environmental obstacle constraints during the tracking process, so that the generated trajectory conforms to the actual situation and automatically responds to changes in the layout of environmental obstacles. The specific implementation method is as follows: Build a constrained adaptive model: Design a device-free Wi-Fi tracking model that incorporates environmental information, takes environmental obstacle features as input, and generates fused features for different environmental features while being tightly integrated with the tracking process; Build an automatic map update mechanism: obtain real-time and accurate environmental information to improve tracking accuracy, maintain and update the input map using grayscale images generated from historical trajectory data, and automatically adapt to the changing environment layout; Perform constraint-enhanced training: select trajectories near obstacles by filtering simulation data to improve the model's ability to learn constraint information; during the training phase, enhance the ability to learn constraints through data filtering and sequential freezing modules.

2. The Wi-Fi tracking method based on environmental obstacle constraint adaptation according to claim 1, characterized in that: The constraint adaptive model includes a PLCR extractor, a map encoder, a fusion module and a decision module; The PLCR extractor is used to quickly retrieve preliminary trajectory features from the PLCR map; The method of obtaining PLCR using PLCR extractor includes the following steps: first, the CSI data is cleaned based on CSI ratio and various filtering techniques; then, short-time Fourier transform is applied to determine the time distribution of DFS and calculate the related PLCR; finally, long short-term memory is used to extract PLCR feature sequence and generate initial trajectory feature vector; The map encoder is used to identify the distribution and position of obstacles in the environment. First, an obstacle map is constructed as an input map, and then an encoder is designed to divide the constructed obstacle map into multiple sub-blocks, and the distribution of 0 pixel values ​​in the image of each sub-block area is extracted respectively to obtain obstacle features; The fusion module is used to integrate the map features into the preliminary trajectory features; the fusion module expands the map features by adding an extra dimension that matches the sequence length in the initial trajectory features, and connects the two along the extra dimension to obtain a combined feature; and further uses a gated recurrent unit network to process the obtained combined feature, and dynamically refines the trajectory features using obstacle information to obtain rich fused features; The decision module is used to map the integrated features to the output trajectory, use a fully connected layer to map the fused features, and generate the final two-dimensional predicted trajectory coordinates; the decision module captures the influence of the fused features on the predicted trajectory coordinates, and helps the constraint adaptive model learn the inaccessible areas caused by obstacles in the output space.

3. The Wi-Fi tracking method based on environmental obstacle constraint adaptation according to claim 2, characterized in that: The map automatic update mechanism described in the present invention specifically includes the following contents: Construct a probability grayscale map: introduce a probability matrix and a grayscale map, and represent the existence of obstacles in the environment area as a probability value. The size of the probability value is proportional to the possibility of the obstacle appearing at that location. Each value in the probability matrix corresponds to the probability of the existence of an obstacle in the real environment. Update the probability matrix: Use the user behavior history to estimate the probability of an obstacle at each point in the area and complete the update of the probability matrix; Update dynamic grayscale map: Use the locally stored probability matrix to dynamically update the locally stored input map, including: 1) Initialize the input map according to environmental conditions; 2) Generate predicted trajectories from the collected user data; 3) Update the probability matrix based on the predicted trajectory to determine whether the input map needs to be updated.

4. The Wi-Fi tracking method based on environmental obstacle constraint adaptation according to claim 3, characterized in that: The goals of the constraint augmentation training are to: 1) Minimize the distance between the predicted trajectory and the target trajectory; 2) Minimize the intersection between the predicted trajectory and the obstacle; The training goal realization process includes the following: The mean square error is used to measure the distance between the predicted trajectory and the target trajectory: Among them, n represents the number of points on the target trajectory, (x i ,y i ) represents a point on the target trajectory, represents a point on the predicted trajectory; Define an additional loss function to quantitatively evaluate the intersection between the predicted trajectory and obstacles to ensure that the loss function also supports gradient propagation; use the F.grid_sample function in the PyTorch framework to query and extract pixel values ​​from the grid specified position; extract pixel values ​​from the input obstacle map corresponding to each predicted trajectory point, and quantify the degree of intersection between the predicted trajectory and the obstacle by calculating the difference between the pixel value and the 1-pixel map, and then obtain the obstacle loss. The specific function is expressed as: Among them, L OC Indicates the obstacle loss, Represents the pixel value in the input obstacle map corresponding to the predicted trajectory point.

5. The Wi-Fi tracking method based on environmental obstacle constraint adaptation according to claim 4, characterized in that: The strategy of the restriction enhancement training is: The PLCR extractor, map encoder, fusion module and decision module are divided into two parts for modular training. The specific contents are as follows: The first part is the tracking network, which includes the PLCR extractor and decision model, and is used to implement basic tracking functions; The second part is the map network, which consists of a map extractor and a fusion module, used to extract environmental information as auxiliary features; The specific training process is as follows: 1) Establish the basic tracking effect of the tracking network; 2) Initialize the map network to ensure its compatibility with the tracking network; 3) screening the training data to select cases where the predicted trajectory intersects with obstacles; 4) Using the screened data, combined with the mean square error and obstacle error, a fusion loss is designed to perform adversarial training on the map network.

Images